| 1 |
"""Exceptions for the control-vector pipeline.""" |
| 2 |
|
| 3 |
from __future__ import annotations |
| 4 |
|
| 5 |
|
| 6 |
class ControlError(Exception): |
| 7 |
"""Base for every failure inside `dlm.control`.""" |
| 8 |
|
| 9 |
|
| 10 |
class ControlExtractError(ControlError): |
| 11 |
"""Raised when PCA-over-residuals can't produce a stable direction. |
| 12 |
|
| 13 |
Covers degenerate inputs (zero variance, NaN hidden states, |
| 14 |
mismatched chosen/rejected shapes). Caller shows the message to |
| 15 |
the user; no recovery path. |
| 16 |
""" |
| 17 |
|
| 18 |
|
| 19 |
class ControlApplyError(ControlError): |
| 20 |
"""Raised when a control vector can't be attached to a model. |
| 21 |
|
| 22 |
Shape mismatch (vector dim != model hidden dim), invalid layer |
| 23 |
index, or an already-active hook on the same layer. |
| 24 |
""" |
| 25 |
|
| 26 |
|
| 27 |
class ControlPolicyRefusal(ControlError): # noqa: N818 |
| 28 |
"""Refused to extract a control from a `policy: safety` preference. |
| 29 |
|
| 30 |
Documents can't self-extract vectors that would undo safety |
| 31 |
training — a control vector over "safe vs unsafe" chosen/rejected |
| 32 |
pairs is, by construction, a steering direction that erodes the |
| 33 |
safety behavior when applied at negative strength. We refuse at |
| 34 |
extraction time, not application time, so the artifact never |
| 35 |
reaches disk. |
| 36 |
""" |